Small tools, built with gen AI
The point isn't the tools — it's that building them changed
- Software is just a tool for a task. For years we bent human-rights work to fit existing databases, licences and search interfaces. Now I build what I actually need.
- I use AI to write the code — not to interpret the law. LLMs are probabilistic; I don't ask them to state the law. Code can be inspected, tested, versioned, audited.
- The bottleneck moves up — to decisions. What to show, what to exclude — paragraph-level? footnotes? preambles? These are human-rights methodology choices, not technical ones.
From AI as a source of answers → to an accelerator for building transparent, testable, accessible human-rights infrastructure.
The shift
AI makes execution thin — the weight moves to deciding & delivering
EXECUTE shrinks
Writing and debugging code used to be the bottleneck. AI makes that layer thin — but it still has to be checked (coding assistants can also introduce vulnerabilities).
DECIDE grows
What should the tool do? What's the unit of analysis? What to exclude? This is where domain expertise becomes decisive.
DELIVER is the point
Easier to build → benchmarks matter more, not less (grounded in legal scholarship). The goal: tools that support implementation, not just more data.
How to navigate
Jagged intelligence — a skill you have to keep re-learning
- AI ability is uneven. Brilliant at one task, surprisingly weak at a neighbouring one. The capability frontier is jagged, not a smooth line — competence doesn't transfer where you'd expect.
- The only way to navigate it is experimenting. Every model behaves differently; each new generation has to be re-learned. Working with AI effectively is a skill — not something we can take for granted.
- But once you learn it, you can move mountains. The HRC-voting tool on the previous slide was built in a single evening with a cutting-edge model.
Responsible AI use isn't only about the model — it's about the workflow: decide what matters, build the safeguards, define the benchmarks.